The real immediate danger (beyond security) seems to be incumbent distortion of markets and monopolistic behavior, and resultant cronyism and regulatory capture. Proposals like this "letter" are far more likely to exacerbate such issues than to help. Rather than "do something" vs. "do nothing," let's try to build frameworks for appropriately applying existing legal principles to this new world. At this point, "watch, analyze, and do very little" seems right.
Also treat them like a regular corporation. If a corporation trains employees to hack into systems and has them try to hack into systems and you can contact them to ask one of their employees to hack for you- what happens to that company (outside of a white hat security outfit)?
I really don't know if copyright law is the right mechanism. Consider: if I memorize a musical score, say, Beethoven's 7th Symphony, then subsequently write an original score that inevitably, but yet inadvertently uses ideas from the 7th am I in violation of copyright law? I don't think most people would answer yes, but in the case of an AI that trains on the 7th, then is prompted to write a score in which it inadvertently uses ideas from the 7th my guess is that most people would consider that a copyright violation.
Another example. Consider how much of modern pop music is largely regurgitation of ideas developed by LIttle Richard, Chuck Berry, and Buddy Holly, etc. in the 1950s. Is it all in violation? Obviously not. My point is that I think we need an entirely new legal framework the definition of which is well beyond my pay grade.
If a company sells tools that can be used to burglarize a house, and you can contact them and ask to buy their tools that can be used to burglarize houses and then burglarize houses with them, what happens to that company? Nothing at all. We already treat AI companies like "regular corporations."
An AI or an AI-powered robot is more like an employee. If you hire out an employee of a company the company should train and monitor it properly and be liable for damage
This still holds true. A temp agency can't be held liable for criminal acts of employees hired out unless the agency is aware of the criminal activity.
More generally, extending wide nets for liability beyond the people who intend to break the law is not usually effective in doing anything but enriching lawyers.
In the case of McDonalds it was McDonalds that was improperly operating the machinery that overheated the coffee, so this is a perfect example of 1st party liability. In the lawnmower cases that was user error, not user malfeasance. If you intentionally ran over someone's foot with your lawnmower, the lawnmower company isn't liable. If you could show that an LLM committed a crime by accident and not at the direction of the user you could probably hold the company that made it liable. It seems perfectly consistent.
The initial version likely revealed the agenda of the groups pushing this. The signatures are just to lend their efforts credibility.
This sounds like an effort to get a few friendly “experts” to testify before Congress and endorse some of their brilliant ideas (no data centers!)
If an economist (or whomever) has some good ideas and recommendations or has done some actual analysis, shouldn’t they present these? That would at least move the debate forward.
"This is despite the fact that lots of people in the AI industry think their inventions are going to destroy jobs."
As a rule, businessmen do not know anything special about macroeconomics and no one should pay much attention to them. And techies have a recent history of getting it spectaculrly wrong by supporting Trump's low immigrtion, high deficits, trade restriction anti-growth agenda.
[New Edit] The DO know things about the effects of micronomic policy, but there the problem is they favor policies favoring existing firms, particularly in theor own sector over potentially existing firms.
It's not a policy issue. It would just be nonstop bureaucratic harassment. In the sorry state this country is currently in the administration in power can arbitrarily attack its enemies with existing regulations.
Thank goodness for macro- we can at least annotate consistently inaccurate predictions with math. 😊. They can’t predict the economy accurately (or the impact of a policy like “stimulus”) and we are going to ask them about technology?
An iterative approach (which is how businesses operate) is better than a fixed policy idea one becomes wedded to.
For iteration to work, though, there needs to be accountability and measurement of outcomes (something econ can do). Unfortunately, government doesn’t do accountability and sticks with lots of bad ideas that don’t work,
Do nothing is an extremely prudent way of taking action, especially in the face of potentially dangerous unknowns. I am a retired firefighter who had extensive training in hazardous materials incident mitigation over a course of almost 30 years. If we could assess and contain a release of volatile hazardous materials by essentially “doing nothing,” that was usually the best course of action. Especially if we were unsure of how those materials would act if we interfered too much or too soon in their process of making their actions known.
I tend to be in the camp of AI will take jobs away eventually (when it's good enough), and also hoping that it does. Because I'd love to move towards a post-scarcity civilization that we don't *have* to work in, just if we want to. -- I want the robots and AIs to do all the work, and then we do whatever the hell we want
After retirement my father spent a decade tutoring math for the children of millionaires. It was a great gig for him, and helped several kids ace the SAT and get into Harvard and Stanford.
Sadly, the economy does not allow this type of gig to scale. Abundance might allow for every smart kid to have a private math tutor! A million new soccer coaches! Guitar lessons for all!
Public art, resurgence of community theatre, universal organic neighborhood gardens! My daughter would love to dedicate her life to supporting communies of rabbit lovers, but she's busy working 50+ hours each week at a bullshit job that pays enough to pay rent.
There are infinite meaningful endeavors that are economically infeasible at scale because of high labor demand for bullshit jobs. Abundance changes the equation.
Post-scarcity means that you can have almost anything physical that you want for basically free because providing it is too cheap to care about. Like how, if you're on an unlimited phone plan, you don't have to count the minutes you spend talking on the phone, or how you don't have to pay to download books that are in the public domain.
Post-scarcity meaning energy, food, housing abundance - think utopia where everyone has all basic needs met because we as humanity can easily provide them.
On the whatever we want looking like a job, yeah probably, or like taking hobbies very seriously, or whatever. I think humans like structure, routine, meaning, and community - but we don't need a job per se and we can get all that from other things that may look job-like.
On top of all of this, which is. great and succinct and informative, there is the reality that *considered inaction is itself an action*. Take that amazing close-up shot that went viral on the internet of Haaland hunting a header for a goal. Much of the time he is "just standing there". But he's not just standing there as though he has suddenly just found himself present on a football pitch — he's avoiding the temptation to perform arbitrary activities that put him in worse shape to do what he might need eventually to do.
Yeah, I can't wait for the retrospective post Noah makes in a few years about what came of this.
---------------------------
AI is a serious advancement. It feels different bc it IS different. Analysts are missing how different it is.
Human technology is built from a bunch of parts that each has to be designed to work together to achieve some task. As the task gets harder (jetliner vs skateboard) the parts are assembled into hierarchical modules, but each module needs to perform its sub-task, and has well defined performance specs and inputs and outputs, and nowadays, its own embedded linear programming firmware. And critical failure paths.
Biology works differently. It is FAR MORE complex and hierarchical than human technology (in terms of the number and types of parts involved), and their specifications and what each part does often appears unclear, random, and/or redundant. And the system is highly robust to removing individual parts or even types of parts. And it was not designed with attention to each piece, it emerged as the result of a long and costly optimization process just to perform an overall task (called natural selection).
LLMs and other deep learning products are the FIIRST commercial human technologies that work from scratch the way biology works: emergent, complex, fuzzy, overparameterized. And they are the result of long and costly optimization processes to perform various tasks.
These deep learning algos are the first of a new _branch_ of human technology that is fundamentally different and more capable than all earlier human technology. One that works the way biology works. Not biomimetic. But embodying biology's secret design sauce!
But AI is just software. It can only perform 'information' tasks. Whereas learning optimization is a design approach. Biology shows us that deep learning can be used to create highly capable and robust materials, devices and autonomous systems embodied in the real world, that can perform very complex real world tasks. That is 'deep learning hardware'.
In a couple (human) generations, our technology hardware will be largely unrecognizable to us. The way it is designed and manufactured will be alien. It will not be designed... it will be optimized (or imagined) on large scale AI platforms. It will be printed/grown with a massive amount of embedded information and part density. And it will be performant and robust it ways that will supercede biology, while looking (to us) uncannily biological.
The policies have already been written by some group and given to a few members of Congress. All they are waiting for is to a shell a few “experts” in for congressional testimony to submit the bill.
Then our elected reps not only get cash from the activists pushing the policies but also cash from tech companies buying protection money and lobbying against them.
It's just corruption with extra steps. Worse than traditional corruption, actually. With our system everyone has to pretend they are actually doing something useful to make it not look like graft, so not only do you have to pay them off, but also you need to hire your own lawyers and bureaucrats to manage the payoff process which can take years. It would actually be much more efficient if you could just hand an envelope to a corrupt official and be on your way.
We must take action to form a committee to decide who will be on the committee that will propose a process by which we will appoint committee to decide on a policy.
Exactly. This obsession with "steering" AI through bureaucratic hand-waving is a classic case of the blind leading the sighted. History shows us that when central planners try to micromanage technological evolution with top-down mandates, it always ends in economic comedy—or tragedy, as with the backyard furnaces Noah pointed to.
Noah's pragmatism is so refreshing. The truth is, the vast majority of society—bureaucrats and elite academics included—will simply adapt to the AI wave, not direct it. Let technology evolve, and let market forces do their job. If entrepreneurs are willing to put their own capital and reputation on the line to take the risk, let them. We don't need a panel of self-appointed mandarins adding noise to an already complex ecosystem.
The one immediate thing that AI is doing: It's rapidly driving up demand for 24/7 base load electricty, right when we're trying to rapidly retire dirty fossil fuel Industrial Tier generation. Now, people have already found credible solutions for that problem: Rapidly construct widely distributed solar parking lot Virtual Power Plants on all of our ubiquitous, ridiculously under-utilized parking lots, everywhere. France and South Korea are already doing it, on ALL lots larger than 80 spaces, nationwide, within 3 to 5 years, largest lots first. While we're wasting time dithering about AI.
Before or in addition to trying to understand the economics of LLMs i think it is important to understand their psychological, sociological and educational effects. Do they reduce cognitive endurance? Do they decrease human interaction? etc. Many people think they know the answers to these questions but I think we need research to get a better understanding. We also need to understand how education needs to change. Understanding must precede action.
This is the part of the debate the economics conversation usually skips past, and it's not a minor gap - cognitive and educational effects run on a different, probably faster, clock than labour-market ones.
Worth knowing there's already a first data point. MIT Media Lab's EEG study of essay-writing found LLM users had the weakest neural connectivity of the three groups tested, the lowest sense of ownership over their own writing, and - months later - measurably worse recall than people who'd written unaided. Early and narrow, but it's a real answer to "does it reduce cognitive endurance," not just a hunch. (https://www.media.mit.edu/publications/your-brain-on-chatgpt/)
Where I'd push back slightly: "understanding must precede action" is clean in principle but risky here, because education is exactly the kind of system where waiting has its own cost - a cohort that goes through school before anyone's worked out the right answer doesn't get a do-over. The choice isn't really act-blind versus wait-for-certainty; it's building in reversibility and measuring as you go, rather than picking one over the other.
I agree with the argument that predicting future jobs is difficult, and it hasn't been reliably done. The novel uses of a novel technology are often surprising, which should be unsurprising. Pretending that you have a crystal ball (like in the radiologist example) is a failing proposition. Thus, steering research to be consistent with your guesses here is going to be hard. How much weight do you give to those guesses?
That said, we shouldn't pretend like there isn't a form of steering already. You might say it's set by the market, though you'd also have to acknowledge the market works in less direct ways here than it does upon things like the price of oranges. The idiosyncrasies of market participants enter in here, and if a significant group of market participants share the same idiosyncrasy that's wrong, it wouldn't get corrected until hard evidence that they're wrong became widely enough known.
I do though think we can make some efforts, rather than just wait and see. For one, we can look back at weaknesses we've overlooked in the past. Since any of these could become more important, and we've managed to get by with overlooking them is due to their effect being minimized, a good use of effort in the early change days is shoring up these potential weaknesses.
My own hobby-horse here is advertising. I think it's already problematic, but we've ignored it. To be clear, it's not a concern driven by a personal annoyance, but one rooted in the non-productive aspects of the enterprise once it's passed from the "informed potential customers" point, and trends into "manipulate possible customers".
I think we also should try to make predictions. Some will be wrong. We shouldn't act decisively on those where we shouldn't be confident. But just the attempt at making them, will allow us to recognize when decisive data starts to emerge. My own attempt here: https://substack.norabble.com/p/ai-jobs-the-hidden-rules-of-demand
If the concern is employment loss, then the policy implication is to facilitate re-allocation of labor from activities where labor is being displaced to other activities. Otherwise, the only way to avoid loss is for wages to fall in the labor dispacing activities. This can happen at all levels from the individual, firm, or economic sector. YIMBY-ism/Abundance is policy to make AI complimentary at the macro level.
The real immediate danger (beyond security) seems to be incumbent distortion of markets and monopolistic behavior, and resultant cronyism and regulatory capture. Proposals like this "letter" are far more likely to exacerbate such issues than to help. Rather than "do something" vs. "do nothing," let's try to build frameworks for appropriately applying existing legal principles to this new world. At this point, "watch, analyze, and do very little" seems right.
Copyright law and data privacy would be a start
Also treat them like a regular corporation. If a corporation trains employees to hack into systems and has them try to hack into systems and you can contact them to ask one of their employees to hack for you- what happens to that company (outside of a white hat security outfit)?
I really don't know if copyright law is the right mechanism. Consider: if I memorize a musical score, say, Beethoven's 7th Symphony, then subsequently write an original score that inevitably, but yet inadvertently uses ideas from the 7th am I in violation of copyright law? I don't think most people would answer yes, but in the case of an AI that trains on the 7th, then is prompted to write a score in which it inadvertently uses ideas from the 7th my guess is that most people would consider that a copyright violation.
Another example. Consider how much of modern pop music is largely regurgitation of ideas developed by LIttle Richard, Chuck Berry, and Buddy Holly, etc. in the 1950s. Is it all in violation? Obviously not. My point is that I think we need an entirely new legal framework the definition of which is well beyond my pay grade.
If a company sells tools that can be used to burglarize a house, and you can contact them and ask to buy their tools that can be used to burglarize houses and then burglarize houses with them, what happens to that company? Nothing at all. We already treat AI companies like "regular corporations."
The word I used is “employee”, not tool.
An AI or an AI-powered robot is more like an employee. If you hire out an employee of a company the company should train and monitor it properly and be liable for damage
This still holds true. A temp agency can't be held liable for criminal acts of employees hired out unless the agency is aware of the criminal activity.
More generally, extending wide nets for liability beyond the people who intend to break the law is not usually effective in doing anything but enriching lawyers.
It doesn’t even hold true for tools and inanimate objects,
Ask any lawnmower companies about multimillion judgements for accidents caused by user error. Ask McDonalds about hot coffee.
There shouldn’t be separate rules or excuses for tech, IMO.
In the case of McDonalds it was McDonalds that was improperly operating the machinery that overheated the coffee, so this is a perfect example of 1st party liability. In the lawnmower cases that was user error, not user malfeasance. If you intentionally ran over someone's foot with your lawnmower, the lawnmower company isn't liable. If you could show that an LLM committed a crime by accident and not at the direction of the user you could probably hold the company that made it liable. It seems perfectly consistent.
Good decision.
The initial version likely revealed the agenda of the groups pushing this. The signatures are just to lend their efforts credibility.
This sounds like an effort to get a few friendly “experts” to testify before Congress and endorse some of their brilliant ideas (no data centers!)
If an economist (or whomever) has some good ideas and recommendations or has done some actual analysis, shouldn’t they present these? That would at least move the debate forward.
"This is despite the fact that lots of people in the AI industry think their inventions are going to destroy jobs."
As a rule, businessmen do not know anything special about macroeconomics and no one should pay much attention to them. And techies have a recent history of getting it spectaculrly wrong by supporting Trump's low immigrtion, high deficits, trade restriction anti-growth agenda.
[New Edit] The DO know things about the effects of micronomic policy, but there the problem is they favor policies favoring existing firms, particularly in theor own sector over potentially existing firms.
Silicon Valley supports none of those things. They accepted them as a compromise because the Democrats have it in for them.
But these are worse than Democratic policies??? How could D's be worse? Becasue Lisa Kahn says mean thngs? :)
It's not a policy issue. It would just be nonstop bureaucratic harassment. In the sorry state this country is currently in the administration in power can arbitrarily attack its enemies with existing regulations.
Thank goodness for macro- we can at least annotate consistently inaccurate predictions with math. 😊. They can’t predict the economy accurately (or the impact of a policy like “stimulus”) and we are going to ask them about technology?
An iterative approach (which is how businesses operate) is better than a fixed policy idea one becomes wedded to.
For iteration to work, though, there needs to be accountability and measurement of outcomes (something econ can do). Unfortunately, government doesn’t do accountability and sticks with lots of bad ideas that don’t work,
Right on Noah!!! Let’s see what “it” is before we regulate it!!
This type of intellectual hubris does nobody no good. It’s holding back Europe and leading the democratic socialists to lord knows where.
But once we figure out what “it” is and what downsides it brings about then we can have a good discussion of what fixes it needs.
The "fixes" are likely to be things we should already be doing.
Or things that will be useless or counterproductive
Do nothing is an extremely prudent way of taking action, especially in the face of potentially dangerous unknowns. I am a retired firefighter who had extensive training in hazardous materials incident mitigation over a course of almost 30 years. If we could assess and contain a release of volatile hazardous materials by essentially “doing nothing,” that was usually the best course of action. Especially if we were unsure of how those materials would act if we interfered too much or too soon in their process of making their actions known.
I tend to be in the camp of AI will take jobs away eventually (when it's good enough), and also hoping that it does. Because I'd love to move towards a post-scarcity civilization that we don't *have* to work in, just if we want to. -- I want the robots and AIs to do all the work, and then we do whatever the hell we want
I think people (especially men) need to work. To feel like they are producing value. To feel like they are providing for their family.
That doesn't mean we need to stick with 40 hours a week.
I'm thinking maybe 6 hours a day 4 a week could work.
Are you sure, or is that just now, under our current system?
Hard to test of course. But I do believe that its generally true.
Anecdote, my grandfather retired at 45. Took up drinking instead.
It's really important to keep busy
But "whatever we want" will probably still look like a job. [What could "post scarcity" mean?]
After retirement my father spent a decade tutoring math for the children of millionaires. It was a great gig for him, and helped several kids ace the SAT and get into Harvard and Stanford.
Sadly, the economy does not allow this type of gig to scale. Abundance might allow for every smart kid to have a private math tutor! A million new soccer coaches! Guitar lessons for all!
Public art, resurgence of community theatre, universal organic neighborhood gardens! My daughter would love to dedicate her life to supporting communies of rabbit lovers, but she's busy working 50+ hours each week at a bullshit job that pays enough to pay rent.
There are infinite meaningful endeavors that are economically infeasible at scale because of high labor demand for bullshit jobs. Abundance changes the equation.
Post-scarcity means that you can have almost anything physical that you want for basically free because providing it is too cheap to care about. Like how, if you're on an unlimited phone plan, you don't have to count the minutes you spend talking on the phone, or how you don't have to pay to download books that are in the public domain.
See you on the Enterprise :)
Pretty much!
Post-scarcity meaning energy, food, housing abundance - think utopia where everyone has all basic needs met because we as humanity can easily provide them.
On the whatever we want looking like a job, yeah probably, or like taking hobbies very seriously, or whatever. I think humans like structure, routine, meaning, and community - but we don't need a job per se and we can get all that from other things that may look job-like.
On top of all of this, which is. great and succinct and informative, there is the reality that *considered inaction is itself an action*. Take that amazing close-up shot that went viral on the internet of Haaland hunting a header for a goal. Much of the time he is "just standing there". But he's not just standing there as though he has suddenly just found himself present on a football pitch — he's avoiding the temptation to perform arbitrary activities that put him in worse shape to do what he might need eventually to do.
Don't just do something, stand there!
Yeah, I can't wait for the retrospective post Noah makes in a few years about what came of this.
---------------------------
AI is a serious advancement. It feels different bc it IS different. Analysts are missing how different it is.
Human technology is built from a bunch of parts that each has to be designed to work together to achieve some task. As the task gets harder (jetliner vs skateboard) the parts are assembled into hierarchical modules, but each module needs to perform its sub-task, and has well defined performance specs and inputs and outputs, and nowadays, its own embedded linear programming firmware. And critical failure paths.
Biology works differently. It is FAR MORE complex and hierarchical than human technology (in terms of the number and types of parts involved), and their specifications and what each part does often appears unclear, random, and/or redundant. And the system is highly robust to removing individual parts or even types of parts. And it was not designed with attention to each piece, it emerged as the result of a long and costly optimization process just to perform an overall task (called natural selection).
LLMs and other deep learning products are the FIIRST commercial human technologies that work from scratch the way biology works: emergent, complex, fuzzy, overparameterized. And they are the result of long and costly optimization processes to perform various tasks.
These deep learning algos are the first of a new _branch_ of human technology that is fundamentally different and more capable than all earlier human technology. One that works the way biology works. Not biomimetic. But embodying biology's secret design sauce!
But AI is just software. It can only perform 'information' tasks. Whereas learning optimization is a design approach. Biology shows us that deep learning can be used to create highly capable and robust materials, devices and autonomous systems embodied in the real world, that can perform very complex real world tasks. That is 'deep learning hardware'.
In a couple (human) generations, our technology hardware will be largely unrecognizable to us. The way it is designed and manufactured will be alien. It will not be designed... it will be optimized (or imagined) on large scale AI platforms. It will be printed/grown with a massive amount of embedded information and part density. And it will be performant and robust it ways that will supercede biology, while looking (to us) uncannily biological.
At this stage, shouldn’t an “act now” statement focus on proposing a process for developing policies rather than policies themselves?
The policies have already been written by some group and given to a few members of Congress. All they are waiting for is to a shell a few “experts” in for congressional testimony to submit the bill.
Then our elected reps not only get cash from the activists pushing the policies but also cash from tech companies buying protection money and lobbying against them.
This is the way the system works.
It's just corruption with extra steps. Worse than traditional corruption, actually. With our system everyone has to pretend they are actually doing something useful to make it not look like graft, so not only do you have to pay them off, but also you need to hire your own lawyers and bureaucrats to manage the payoff process which can take years. It would actually be much more efficient if you could just hand an envelope to a corrupt official and be on your way.
We must take action to form a committee to decide who will be on the committee that will propose a process by which we will appoint committee to decide on a policy.
...and by creating all that bureaucracy, we've already solved the problem of keeping people employed! ;-)
Yeah: or: process before hot takes?
This is just more Doom Marketing.
Exactly. This obsession with "steering" AI through bureaucratic hand-waving is a classic case of the blind leading the sighted. History shows us that when central planners try to micromanage technological evolution with top-down mandates, it always ends in economic comedy—or tragedy, as with the backyard furnaces Noah pointed to.
Noah's pragmatism is so refreshing. The truth is, the vast majority of society—bureaucrats and elite academics included—will simply adapt to the AI wave, not direct it. Let technology evolve, and let market forces do their job. If entrepreneurs are willing to put their own capital and reputation on the line to take the risk, let them. We don't need a panel of self-appointed mandarins adding noise to an already complex ecosystem.
In all the AI discourse, does anyone else ever hear Donald Fagen's 1982 song "IGY" in your head?
A just machine to make big decisions
Programmed by fellows with compassion and vision
We'll be clean when their work is done
We'll be eternally free, yes, and eternally young
(although my favorite line is still "There'll be spandex jackets, one for everyone.")
The one immediate thing that AI is doing: It's rapidly driving up demand for 24/7 base load electricty, right when we're trying to rapidly retire dirty fossil fuel Industrial Tier generation. Now, people have already found credible solutions for that problem: Rapidly construct widely distributed solar parking lot Virtual Power Plants on all of our ubiquitous, ridiculously under-utilized parking lots, everywhere. France and South Korea are already doing it, on ALL lots larger than 80 spaces, nationwide, within 3 to 5 years, largest lots first. While we're wasting time dithering about AI.
Before or in addition to trying to understand the economics of LLMs i think it is important to understand their psychological, sociological and educational effects. Do they reduce cognitive endurance? Do they decrease human interaction? etc. Many people think they know the answers to these questions but I think we need research to get a better understanding. We also need to understand how education needs to change. Understanding must precede action.
This is the part of the debate the economics conversation usually skips past, and it's not a minor gap - cognitive and educational effects run on a different, probably faster, clock than labour-market ones.
Worth knowing there's already a first data point. MIT Media Lab's EEG study of essay-writing found LLM users had the weakest neural connectivity of the three groups tested, the lowest sense of ownership over their own writing, and - months later - measurably worse recall than people who'd written unaided. Early and narrow, but it's a real answer to "does it reduce cognitive endurance," not just a hunch. (https://www.media.mit.edu/publications/your-brain-on-chatgpt/)
Where I'd push back slightly: "understanding must precede action" is clean in principle but risky here, because education is exactly the kind of system where waiting has its own cost - a cohort that goes through school before anyone's worked out the right answer doesn't get a do-over. The choice isn't really act-blind versus wait-for-certainty; it's building in reversibility and measuring as you go, rather than picking one over the other.
I agree with the argument that predicting future jobs is difficult, and it hasn't been reliably done. The novel uses of a novel technology are often surprising, which should be unsurprising. Pretending that you have a crystal ball (like in the radiologist example) is a failing proposition. Thus, steering research to be consistent with your guesses here is going to be hard. How much weight do you give to those guesses?
That said, we shouldn't pretend like there isn't a form of steering already. You might say it's set by the market, though you'd also have to acknowledge the market works in less direct ways here than it does upon things like the price of oranges. The idiosyncrasies of market participants enter in here, and if a significant group of market participants share the same idiosyncrasy that's wrong, it wouldn't get corrected until hard evidence that they're wrong became widely enough known.
I do though think we can make some efforts, rather than just wait and see. For one, we can look back at weaknesses we've overlooked in the past. Since any of these could become more important, and we've managed to get by with overlooking them is due to their effect being minimized, a good use of effort in the early change days is shoring up these potential weaknesses.
My own hobby-horse here is advertising. I think it's already problematic, but we've ignored it. To be clear, it's not a concern driven by a personal annoyance, but one rooted in the non-productive aspects of the enterprise once it's passed from the "informed potential customers" point, and trends into "manipulate possible customers".
I think we also should try to make predictions. Some will be wrong. We shouldn't act decisively on those where we shouldn't be confident. But just the attempt at making them, will allow us to recognize when decisive data starts to emerge. My own attempt here: https://substack.norabble.com/p/ai-jobs-the-hidden-rules-of-demand
If the concern is employment loss, then the policy implication is to facilitate re-allocation of labor from activities where labor is being displaced to other activities. Otherwise, the only way to avoid loss is for wages to fall in the labor dispacing activities. This can happen at all levels from the individual, firm, or economic sector. YIMBY-ism/Abundance is policy to make AI complimentary at the macro level.